Hacker News Reader

Comment by roywiggins | Hacker News Reader

the thing about cells is that even labeling bits of them as discrete "parts" seems like it might be misleading, eg this from the linked paper:

> recent research on the cellular architecture demands that we look more carefully at what we have previously assumed were well-defined structures and reconsider them as stabilized processes. Because processes are temporally extended, it follows that they can only be understood by giving time due consideration... The structure of a machine, after all, can be grasped in abstraction from time (as it is not constantly changing), whereas the structure of, say, a whirlpool or a stream cannot. This explains why, when we have started using techniques that allow us to examine the cellular architecture in real time, we have found that many of the cell’s compartments and organelles are not fixed machineries at all, but stable macromolecular fluxes.

View on HN

Okay, „parts“ for me was just a word to refer to the machine, it is also okay if it is only a single part. It still remains a machine in the sense that it performs a task it should perform and does so using the same approach across all instances.

That the „parts“ do not only perform a specific function is not really surprising, because the machine does not need to be comprehensible or understandable. I would expect single-purpose parts only if that helps the function or if it helps evolvability (in the sense of not creating dead ends).

The paper covers this also. Cells don't behave predictably:

> The variable flickering of transcriptional activity in different cells is one of the major causes of heterogeneity in isogenic populations. But where exactly does this cell-to-cell variability in transcriptional activity come from? The answer becomes apparent when we remember that gene expression is a molecular process, and like all molecular processes, it is inherently stochastic, given that it takes place in an environment that is subject to the chaotic dynamics of Brownian motion. Each step in the process relies on fortuitous encounters between molecules that are randomly moving about as a consequence of thermal agitation. Evidently, these molecules must be at the right place and at the right time—not to mention in the right vibrational state—for them to be able to participate in the appropriate reactions. The unpredictability of the whole process is further amplified by the fact that the participating molecules in each step are present in the cell in very low copy-numbers, as this decreases the chances of successful interactions between them...

> as research into the non-genetic heterogeneity of cells continues, evidence for the biological importance of this phenomenon is mounting ever-rapidly. We now know that non-genetic heterogeneity plays key roles in both microbial and eukaryotic cells, in embryonic development, and in evolution. For one thing, it is a crucial generator of phenotypic diversity, which enables cell populations to adapt rapidly to changing environmental conditions. It does so by permitting the implementation of probabilistic diversification strategies within a population, such as bet-hedging and divisions of labour, which can confer considerable fitness advantages...

> One very important theoretical implication of the probabilistic nature of cellular behaviour and the observed heterogeneity of cell populations is that, quite literally, every cell (in an organism and elsewhere) is a unique entity. No two cells are identical, given that no two cells respond to a stimulus in the exact same way—even if they are genetically the same.

It's just a very, very optimized & elegant design. Reminds me of this:

     He had located the data he was working on
     near the top of memory ---
     the largest locations the instructions could address ---
     so, after the last datum was handled,
     incrementing the instruction address
     would make it overflow.
     The carry would add one to the
     operation code, changing it to the next one in the instruction set:
     a jump instruction.
     Sure enough, the next program instruction was
     in address location zero,
     and the program went happily on its way.

And it's not like our own transistors aren't probabalistic when you look closely enough.

Anecdote: had an 8086 system-call jump table that fielded up to 64K calls, but using a jump table of 4K. Address mode was segment:offset where segment was 17 bits and offset 16, but shifted by 4. The table consisted of the same trap instruction throughout, which took 4 bytes. If I remember it all right.

Anyway the system call trap examined the caller stack, found the return address (to the trap table), picked out the combination of segment and offset used to construct a system-call index. Following so far?

We went paged. Changed syscall segment to indirect, 16 segment descriptors all mapped to the start of the jump table. Started failing on one particular system call out of thousands - it would page fault! Why?

The auto-increment and prefetch of the trap opcode in the jump table would fault, for only one syscall, the one that mapped to the last entry in the trap table where the offset came out to 0FFC. Wrapped to 1000, which with that particular segment meant next-page! Which was unmapped (system trap table was in low memory, alone).

My solution: change the syscall index for that operation to something else, and document in the syscall comments to never use it.

If a computer running a probabilistic calculation got it right 90% of the time, and the next one over running the same calculation on identical hardware got it right 10% of the time, there would be something wrong with the computer. This is not true with cells.

Your point only stands when the percentages are exaggerated. What if it were 99.99 vs 99.999? Let's not forget cosmic ray bitflips. Just because the two machines operate at different levels of probabalistic exactitiude, does not mean they are not both machines.

Do you really think modern computers are fully deterministic at the lowest levels?

Cells behaving randomly is a fundamental part of their ordinary operation. Computers are tractable because we've done our best to engineer out the randomness and, when we can't, layer on enough error correcting codes so we can mostly not think about it.

But cells don't operate on such nicely bifurcated scales (small and random vs large and mostly predictable). If they did, the dream of cellular wiring diagrams allowing cells to be understood would be a lot more achievable.

Your example of randomness upthread was brownian motion: I'm not surprised a transport mechanism relies on stochastic behaviors. To me, that looks comparable to electrons bouncing down a wire, or minimizing quantum tunneling effects between transistor layers - random at a very low level, but averages out to the behavior we need. Put guardrails at the extremes, so the random behavior stays within bounds, within a margin of error - that's just good engineering.

Compare to loggers, floating trimmed trees down a river to the sawmill: the logs bouncing down the river are behaving in an entirely random manner, but it's within bounds and doesn't matter, because they'll still make it downstream where they get caught by another system. It's a "machine" with random behaviors. Maybe this (obligatory) xkcd illustrates my point: https://www.explainxkcd.com/wiki/index.php/2916:_Machine

I get the impression that what this is coming down to is massive complexity rendering the machine incomprehensible - but again, that doesn't mean it isn't a machine.

If sending logs down a river is a machine, couldn't that make almost anything a machine? The river itself, for instance.

Yes, it is a mechanical system that invokes stochastic processes within deterministic bounds to accomplish a specific, complex outcome. Given sufficient knowledge, it can be mapped and its functions defined.

I believe that summarizes transistors, cells, and forest-to-sawmill machines.

Doesn't that also describe, like, plate tectonics and tornadoes? Once you expand the definition of "machine" that far, what doesn't it reach?

Vortexes absolutely have their place in fluid dynamics and mechanical design. Modern shipbuilding depends on them. If we ever manage to attempt geoengineering at the largest scale, plate tectonics would probably also come into play.

Which is to say, yes: it is a system composed of comprehensible systems, which makes it, itself, a comprehensible system. And that's the whole point here: given knowledge and time, we can tease out exactly how it works.

Are you saying a, for example, muscle cell that produced 10% of the contracting power of its neighbor for the same sugar cost would be perfectly healthy?

It's not that cells can't go awry, it's that perfectly healthy, otherwise identical, cells do not behave predictably to identical stimuli, even on average. That's their normal mode of existence. Some of those behaviors are pathological (eg cancer) but mostly they're not.

Computers don't behave this way at all: you can be reliably sure that if two computers produce different results with the same input that one of them is wrong. If it's a stochastic algorithm at least the averages should be within a certain bound. You can't make the same conclusion with cells.

Yes you can and the comment you responded to illustrated this very well.

Which stochastic algorithms produce different distributions of results on different identical machines with the same inputs, but it's not considered a bug?

> different identical machines with the same inputs

> No two cells are identical, given that no two cells respond to a stimulus in the exact same way—even if they are genetically the same.

That's just begging the question.

> remember that gene expression is a molecular process, and like all molecular processes, it is inherently stochastic, given that it takes place in an environment that is subject to the chaotic dynamics of Brownian motion

Wait until you hear about quantum effects in microprocessors!

I think when somebody says "machine", a lot of people just think victorian steam engines. Machines can cover a lot more. In fact in the future we will build devices that do work similar to what's going on in the cells, and these devices will have to obey the same laws of physics - they too will wiggle and randomly do weird things, but on average, they will do what we intend for them to do.

A machine can be arbitrarily complicated.

I read that bit of the paper is saying that genetically identical cells sitting in identical environments don't even behave the same as each other on average.

We designed computers to manage the quantum and thermal noise "at the bottom" so we can mostly ignore it. We go to incredible lengths to make the constituent parts of our machines predictable, and when they aren't, we layer error correction on top. Cells usually don't do that: noise is inherently part of their behavior and isn't suppressed nearly as much as was often assumed.

This is simply not true. If a cell would not behave as intended on average it would stop being a cell.

The fact is that an amoeba cell behaves in a certain way for all amoebas. It does not go around and behave like a stone or like a drop of olive oil or even like a cell of some random bacterium.

It has very clear and obvious functions and the mechanisms inside the cell are the reason it consistently shows amoeba behavior.

If all you need for something to be a machine is "it has long-term stable behaviors and constituent parts that cause that behavior" then it seems like that sweeps in a lot of stuff that manifestly aren't machines, like tornadoes and similar. My gently cooling coffee is behaving like it does because of its constituent parts, but it's not a machine either.

parent was pointing out that cells don't just behave like an improbability drive, they do actually follow certain rules on average, in contrast to grand parent who asserted incorrectly that they don't. if they didn't, they wouldn't be a cell, they would be a soup.

That's kind of the thesis of the paper, that they're much more like highly organized soup than wiggly machines.

> we layer error correction on top. Cells usually don't do that:

I'd hate to break it to DNA correction machinery.

The paper does not cover this.

I didn‘t say every single molecule in the cell acts predictably. I said that the cell „performs a task it should perform and does so using the same approach across all instances.“

It's saying that genetically identical cells can react differently to exactly the same inputs. Some cells are more reactive to particular stimuli than others, even though they have the same environment and same DNA.

That doesn't seem like they're "using the same approach over all instances" to me.

Reply on news.ycombinator.com